What Does Motion Reveal About Transparency ? Moshe Ben-Ezra and Shree K. Nayar Columbia University ICCV Conference October 2003, Nice, France This work was supported by an NSF ITR Award IIS
Transparency is Very Challenging Existence of a transparent object. Finding its shape and pose
Real and Virtual Features Lambertian V1V1 V2V2 F` V2V2 V1V1 F Specular F F`F` V1V1 V2V2 Transparent
Environmental Matting* * Zongker, el al. SIGGRAPH 99, Alternating pattern Object Camera Does not recover shape and pose. Requires controlled environment.
Shape from Polarization in Highlight* * Saito et al. CVPR’99. Object Camera Light Rotating Polarizer Limited to a single interface at the object’s surface. Requires controlled environment. N
Shape from Refraction and Motion* * H. Murase. PAMI, 1992 Camera Water Single interface only. Fixed Pattern
Motion is Key to Transparency
Transparent Shape From Motion Given: Views And a Parametric Model (such as super-ellipse) Recover: Shape: Values of parameters (e, n) Pose: Rotation R, Translation T General analytic solution does not exist.
Transparency From Motion Reversed rays are parallel to each other regardless of the complexity of their paths Distant feature
Approach: Initialization Image Plane
Approach: Initial Guess
Approach: Refine
Error Function (0,0,1) r 1,1.. r 1,n r 2,1.. r 2,n - Object’s shape parameter vector R,T - Object’s pose
Simulation Setup Parallel rays from features Transparent object Camera side rays
Example (Simulation) Single Parameter. Newton-Raphson optimization Initial Guess Symmetric Superellipse (n=e)
Evaluation (Simulation) GTGT Both init Pos res Sphere Ground Truth Initial Guess Computed Result Shape Error GTBoth InitBoth Res Lens GTBoth InitBoth Res Cube GTBoth Init Both Res Water Pipe
Real Experiment: Sphere
Features
Initial Guess
Setup: Initial Guess Initial Guess: Diameter: 8 inch
Setup: Result Ground Truth: Diameter: 3 inch. Computed: 3.18 inch
Result
Real Test: Water Filled Pipe
Features
Initial Guess
Setup: Initial Guess Initial guess: Diameter: 200.0mm Thickness: 20.0mm
Setup: Result Ground Truth: Diameter: 117.0mm Thickness: 3.0mm Computed: Diameter: 116.1mm Thickness: 2.3mm
Result
Real Test: Superquadric
Features
Initial Guess
Result Ground truth: e = ?Computed: e = 0.18
Summary Shape and pose parameters Multiple interfaces No Segmentation required
Parameterizations of Interest Polynomials: modeling surfaces, lenses CAD models: shape of industrial objects Dynamic models: time dependent parameters
Assumptions Camera parameters are known. Features are far* and are trackable. A proper model and a hypothesis (an initial guess) are given. * One possible assumption.
Real Tests Setup
Implementation Features were manually selected and tracked (9 views). Captured rays, a model, refraction index and a hypothesis were given as inputs. Shape and pose were recovered using simple gradient decent (with derivatives).
The Physics of Transparency First Interface: μ 1 → μ 2 Second Interface: μ 2 →μ 1 33 11 11 33 N1N1 N2N2 22 22
Parametric Shape Examples Super-Ellipse 2 parameters Spherical Harmonics 8 parameters No analytic solution